AI Models
Model releases, benchmarks, inference and fine-tuning.
Algorithmic Velocity, Compute Economics, and AI Developer Tools in 2025–2026
Comprehensive analysis of Henry Josephson's Epoch AI research on algorithmic velocity, compute economics, AI models, and agentic workflows in 2025-2026.
Twenty Percent Conversion and Nobody's Optimizing for It
LLM referral traffic converts at 20% — 61% higher than paid search. But the visitor who clicks from a ChatGPT answer needs a completely different landing experience than one who clicks from an ad. Here's how to build for them.
The Cognition Gap: What Three Studies Reveal About Why Machines Think Nothing Like Us
Three experiments — infants versus neural networks, AI on abstract reasoning puzzles, and translators working alongside AI — expose a fundamental asymmetry between human and machine cognition.
Lightricks Spinoff LTX Releases LTX-2.5: What It Means for AI Cloud Infrastructure Companies in India
A grounded look at LTX-2.5’s open weights, video-generation performance, deployment options and implications for AI infrastructure planning in India.
Navigating AI's Dual Current: Using the Daoist Concept of Zhēn as a Benchmark for Wisdom
Exploring how AI simultaneously illuminates and flattens human cognition, using Peter H. Kahn Jr.'s Daoist-inspired concept of Zhēn (真) as a benchmark for wisdom and critical engagement.
The Economics of Leverage: Why Training Frontier AI Models Justifies Massive Upfront Capital
An examination of the economic implications of the 'train-once-deploy-many' property of AI models, comparing AI training scaling dynamics to human labor and exploring why AI labs can justify vastly larger upfront resource expenditures.
Kimi K3 Is the Largest Open-Source Model Yet — and a Signal for AI Cloud Infrastructure Companies in India
Moonshot AI's Kimi K3, a 2.8-trillion-parameter open-weight release, is the largest open-source model yet. What it means for self-hosted inference and for AI cloud infrastructure companies in India.
Beyond Rankings: Building an Enterprise Scorecard for AI Search
Expanded research article exploring enterprise SEO measurement of AI Overviews, AI Mode, and LLM visibility based on Tom Capper's webinar insights and Ahrefs research.
FLUX 3 Action's Small-Model Bet and the Elusive AI General Intelligence Definition
Black Forest Labs claims its 7-billion-parameter open-weight World Action Model tops NVIDIA's RoboLab-120 benchmark while flying drones and beating Doom. Here's what that actually tells us about embodied agents and where the field still falls short.
Why DeepSeek's Janus-Pro-7B Decoupled Its Eyes and Still Won
A technical deep-dive into Janus-Pro-7B's decoupled visual encoder architecture, benchmark dominance, and open-science impact.
Why Abacus AI’s Liberated Qwen Matters for AI Language Models in VS Code
What Abacus AI’s open-weight Liberated-Qwen1.5-72B teaches about instruction hierarchy, model evaluation, and safer developer workflows with AI language models in VS Code.
Keeping Open AI Models Usable: The Case for Preserving the Master Weights
A practical look at how open-source AI and open science depend not only on access, but on preservation, verifiability, and usable paths from model artifacts to real-world adoption.